Design and feasibility of spatially aggregated transform based on compression of hyperspectral images

Abstract

Remote sensing through the use of hyperspectral image sensors is becoming more prevalent. These hyperspectral images consist of large amounts of data, much of which is redundant and must be compressed. The purpose of this thesis was -to investigate and optimize a Karhunen--Loe��ve transform (KLT) based algorithm to compress hyperspectral images. The thesis focused on combining a multi-level Karhunen--Loe��ve transform with spatial aggregation in the form of K-means and JPEG2000 to improve compression performance as compared to similar lossy to lossless compression methods. The developed algorithm is dubbed Spatially Aggregated Multilevel Clustering (SAMLC) KLT. Performance optimizations consisted of finding the best structure for the multi-level Karhunen--Loe��ve transform, number of aggregated groups in the K-means step, as well as general overhead reduction. A host processor independent hardware implementation of the k-means aggregation was also implemented in an FPGA as proof of concept for portable applications. The results of this study concluded that the SAMLC algorithm is competitive with similar methods and has superior lossy compression performance. Additionally, the K-means hardware implementation proved successful. The gains achieved are important for hyperspectral imagery to be able to store or transmit large hyperspectral images in typical remote sensing applications in an efficient manner.